Value of Information Analysis for Complex Simulator Models: Application to Wind Farm Maintenance

Abstract

Many domains require the use of sophisticated simulators to adequately model the effect of chosen alternatives on the decision maker’s value. Decision support in such complex systems brings unique challenges around efficiency, because simulating each combination of inputs can be time-consuming. In this paper, we conduct a value of information (VOI) analysis to study whether one should purchase information about critical input uncertainties in such complex systems. We propose a novel computational approach, where Gaussian processes are used to model the decision maker’s profit as a function of different alternatives and uncertainties. Under this modeling assumption, the expected improvement of the profit is analytically available, which we use to approximate the VOI effectively over batches of simulations, thus avoiding too many computer-intensive evaluations of the system. We illustrate the proposed computational approach with an offshore wind farm maintenance application, where the decision maker relies on outputs from large-scale simulations to determine the optimal vessel fleet mix and number of personnel for operation and maintenance. Such computer-intensive simulations mimic long-term energy production under different input conditions. It is often not possible to explore all the alternatives exhaustively, and one must therefore guide the simulations to run on promising alternatives. We conduct a VOI analysis to study whether one should purchase information about failure rates of wind turbines for this application. The proposed methodologies are general and apply to other domains involving expensive simulators.

A value of information methodology for multiobjective decisions in quantitative set‐based design

“…Other notable research applies single objective VOI analysis with Gaussian processes to evaluate and optimize complex simulation models, 24 as well as Bayesian optimization in developing simulation and scenario libraries for autonomous systems. 25 Using variations of the methods shown in Table 1, along with VOI applications in portfolio decision analysis 28 and project management, 29 we develop a methodology addressing the above limitations.…”

Which predictive uncertainty to resolve? Value of information sensitivity analysis for environmental decision models

“…Multi-criteria decision approaches (MCDA/MCDM) have been used with VoI analysis before (e.g., Bates et al, 2014;Runge et al, 2011;Eidsvik et al, 2015), but mostly have addressed these aspects separately (see Bolam et al, 2019). VoI analysis for continuous probability distribution remains conceptually straightforward, but can become challenging for (conditionally) dependent probability distributions (e.g., Eidsvik et al, 2015;Myklebust et al, 2020). Second, multiple, conflicting perspectives on an environmental decision problem commonly exist.…”

Exploratory-Phase-Free Estimation of GP Hyperparameters in Sequential Design Methods—At the Example of Bayesian Inverse Problems

“…Methods for Sequential Design of Experiments (SDoE) exist for a variety of problems, such as optimization (called Bayesian optimization ) (Kushner, 1964 ; Jones et al, 1998 ; Williams et al, 2000 ; Mockus, 2012 ; Shahriari et al, 2016 ; Frazier, 2018 ), contour estimation (Ranjan et al, 2008 ; Picheny et al, 2010 ), estimation of failure probabilities (Bichon et al, 2008 ; Bect et al, 2012 ; Balesdent et al, 2013 ), value of information analysis (Myklebust et al, 2020 ), and Bayesian inverse problems (Sinsbeck and Nowak, 2017 ; Damblin et al, 2018 ; Teckentrup, 2019 ). In all of these methods, the expensive-to-evaluate model function is described by a Gaussian process emulator (GPE).…”